The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run to execute Python scripts that it generates, specifically experiment.py and plot.py.
Evidence: references/code-patterns.md defines a run_experiment function using subprocess.run(["python", "experiment.py", ...]) to execute the generated code.
[DYNAMIC_EXECUTION]: The skill follows a pattern of generating code variants, reflecting on their performance, and executing them in an iterative loop to improve results.
Evidence: SKILL.md defines actions for generate, improve, and debug. references/code-patterns.md includes patterns for Hill-Climbing Code Optimization and Experiment Execution Loops that execute code generated at runtime.
[INDIRECT_PROMPT_INJECTION]: The skill accepts research plans and idea descriptions as input, which are used to generate the logic for the executed scripts. This provides a surface where malicious instructions could be embedded in the data to influence the generated code's behavior.
Ingestion points: The $1 research plan/idea description input in SKILL.md is ingested into the agent context.
Boundary markers: No explicit boundary markers or instructions to ignore embedded instructions are defined for the input interpolation.
Capability inventory: The skill has the ability to write files to the project directory and execute shell commands via subprocess.
Sanitization: No sanitization or validation of the input strings is mentioned before they are interpolated into the code generation prompts.

experiment-code